Evolutionary Algorithms and Agricultural Systems

109 Pages·2002·4.965 MB·English

by  David G. Mayer (auth.)

EVOLUTIONARY
ALGORITHMS AND
AGRICULT URAL
SYSTEMS
THE KLUWER INTERNATIONAL SERIES
IN ENGINEERING AND COMPUTER SCIENCE
EVOLUTIONARY
ALGORITHMS AND
AGRICULT URAL
SYSTEMS
by
David G. Mayer
Queensland Beef Industry Institute, Australia
SPR1NGER SCIENCE+BUSINESS MEDIA, LLC
ISBN 978-1-4613-5693-6 ISBN 978-1-4615-1717-7 (eBook)
DOI 10.1007/978-1-4615-1717-7
Library ofCongress Cataloging-in-Publication Data
A C.I.P. Catalogue record for this book is available
from the Library of Congress.
Copyright © 2002 by Springer Science+Business Media New York
Originally published by Kluwer Academic Publishers in 2002
Softcover reprint of the hardcover 1s t edition 2002
AII rights reserved. No part of this publication may be reproduced, stored in a retrieval
system or transmitted in any form or by any means, mechanical, photo-copying, record ing, or
otherwise, without the prior written permission of the publisher,
Springer Science+Business Media, LLC.
Printed on acid-free paper.
The Publisher offers discounts on this book for course use and bulk purchases.
For fUr/her information, send email to<[email protected]>
Contents
Preface VB
Acknowledgments ix
RATIONALE FOR SYSTEMS MODELLING
AGRICULTURAL SYSTEMS MODELS 9
APPLICATION OF EVOLUTIONARY ALGORITHMS TO MODELS 19
APPLICATIONS OF ALTERNATE OPTIMISATION TECHNIQUES 33
COMPARISONS OF OPTIMISATION TECHNIQUES 53
ROBUST PARAMETERS FOR EVOLUTIONARY ALGORITHMS 61
FUTURE DEVELOPMENTS 77
Appendix 1 79
References 89
Index 105
Preface
Systems research is increasingly being used to investigate and analyse a
wide range of real-world problems, including agricultural production systems.
Given a valid, verified model of a particular system, optimisation is a logical
complement to the modelling exercise. Usually, this takes the form of
maximisation of some measure of the system's performance, such as total
production or economic gross margin. This book deals with the practical
application of optimisation techniques, particularly evolutionary algorithms, to
the study and management of these agricultural systems. It should prove useful
to practitioners applying these methods to the optimisation of agricultural or
natural systems, and would also be suited as a text for systems management,
applied modelling, or operations research university subjects. Basic knowledge
in systems research, along with some computing and programming skills, are
assumed.
Models of agricultural systems range widely on both temporal and spatial
scales. Farm-level systems have typically been investigated, but models also
range out to regional, industry and national scales. Short-term (within-year)
profitability and cash-flow issues are common, but the time-frame can be
extended to a hundred years or more, to investigate sustainability and long-term
effects. In addition to the 'direct' economic maximisation of agricultural
systems, optimisation methods have also seen use in the calibration of internal
model parameters to observed data, maximising the rate of genetic gain in
livestock, in agricultural allocation and scheduling problems, and in the
analysis of sustainability issues in natural systems management.
Agricultural models present a number of difficulties with regard to
optimisation. These problems include complex relationships which are not
conducive to the simpler forms of economic modelling (such as linear
programming); biological variability, which usually requires a stochastic
viii
model; the identification of suitable variables to optimise; the high degree of
complexity in these systems, which translates to high dimensionality of the
search-space; frequent interactions between the effects of the various
(assumedly independent) management options; cliffs and discontinuities in the
search-space (where the system is over-utilised, and 'crashes' both biologically
and economically); and the presence of multiple local optima, caused by very
different combinations of management options having similar economic
outcomes.
Any selected optimisation method is required to deal with all these
problems, and reliably return the solution for the global optimum (or a value
suitably close to this). Generally, evolutionary algorithms (including genetic
algorithms, evolution strategies, and hybrid methods) have proven superior
for this task. Depending on the algorithm and the type and usage of the
model, some problems do remain, however evolutionary algorithms contain
a number of advantageous features which largely circumvent these. For the
alternate optimisation techniques (including gradient and direct-search
methods, simulated annealing, and the tabu search strategy), these
difficulties often prove insurmountable. Published studies on the application
of all these methods to agricultural systems are contrasted and compared, in
terms of quality of the final solution and rates of convergence.
Finally, the listed applications are drawn together into an overview, and
the more successful genetic algorithm methodologies and parameters are
discussed. This identifies combinations which are likely to provide robust
performance, given any similar future system being investigated. Directions
of profitable future research are also outlined.
Acknowledgments
This research was conducted whilst I was an employee of the Queensland
Department of Primary Industries, and I gratefully acknowledge the support of
this organisation, along with the encouragement of my supervisors, Drs Pat
Pepper, Peter O'Rourke and Tony Swain.
Sincere thanks are also due to quite a number of scientists for assistance,
namely
• Professors John Belward and Kevin Burrage of the Centre for Industrial
and Applied Mathematics and Parallel Computing, University of
Queensland, for their guidance and expertise with optimisation techniques.
• Personnel in the Climate Impacts and Natural Resource Systems section,
within the Queensland Department of Natural Resources and Mines. In
particular, Greg McKeon, Grant Stone, Wayne Hall, Samantha Watson and
Neil Flood all provided valuable assistance with the development of these
agricultural models.
• The practitioners who have made their optimisation algorithms available
to the general community on the world-wide web, namely John
Grefenstette (GENESIS), Lester Ingber (ASA) and Henrik Widell
(GENIAL).
• Colleagues in the Modelling and Simulation Society of Australia and
New Zealand who assisted with discussions and critiques, in particular
David White, Mark Howden, Chris Dake and Robert Sherlock.
Finally, I give special thanks to my wife Ailsa, and children Nicole, Kerryn,
Janice, Gregory and Cristie for their support and understanding.
Chapter 1
RATIONALE FOR SYSTEMS MODELLING
This introductory chapter outlines the documented benefits of the
systems research approach, and the steps and methodologies typically used
here. Potential discrepancies between the modelled and real-world systems
are discussed, along with interpretational issues. The various types of
models used for the study and optimisation of agricultural systems are
outlined. These include the widely-used and much published linear and
mathematical programming methods, which however do tend to be
constrained representations of the real world, and can give poor results. As
an alternative approach, general (unstructured) simulation models are
promoted as a more flexible and realistic method of representing the system
being modelled.
1. INTRODUCTION
With the advent of more powerful computers, the science or art of
simulation modelling has become more commonplace. Simulation of a
system is the construction and operation of a model which is a valid
representation of the system. Physical models have long been used to
investigate a variety of problems. For example, the effectiveness of design
in boats and aircraft has been tested by the behaviour of smaller-scale
models in tanks or wind-tunnels. Also, erosion and sedimentation models of
proposed harbour and canal developments are commonly used, with some of
these physical models covering hundreds of square metres. Despite the cost
involved, it is obviously advantageous to gain an understanding of potential
problems and solutions prior to spending millions of dollars on such a
development.
D. G. Mayer, Evolutionary Algorithms and Agricultural Systems
© Kluwer Academic Publishers 2002
2 EVOLUTIONARY ALGORITHMS & AGRICULTURAL SYSTEMS
Computer simulation models of agricultural systems are essentially the
same. They are a representation or abstract of reality, expressed in
mathematical and logical terms. As such, they can never fully represent reality,
and must be imperfect. The only perfect model is reality itself. For example,
the best pasture production model, taking account of rainfall, runoff, through
drainage, water use efficiency, humidity, solar radiation, soil fertility, mobile
nutrients, inter- and intra-species competition, etc. will still not produce
accurate results if excessive rainfall produces a flood which submerges the
pasture for a period of time (unless the modeller has allowed for this
occurrence). Despite these shortcomings, it is the intention of most modellers
to construct a model which will simulate reality as well as is possible, and
which may be assumed valid under most circumstances. In particular, the
model should work well in situations similar to those where the modelled
results will be extended to and used in the real-world system.
The logistics and mechanics of model construction are complex, covering
problem definition; construction of systems diagrams; selection and estimation
of key parameters, pathways and relationships; data availability; programming
approach; hardware and software requirements; and verification and validation.
Figure 1 represents a typical systems diagram (from our series of simulation
studies, Mayer 2000) of an agricultural system, from which a systems model
can, and has, been built. These systems research methodologies have been well
described in a range of introductory and advanced texts (Dent and Blackie
1979, Law and Kelton 1982, Bratley et al. 1987, Kleijnen 1987, Ripley 1987).
It is not the intention here to review this diverse field, but rather to consider
applications of agricultural models from a practical point of view.
The development and proving of a systems model can be lengthy and
expensive, with no guarantee of success, so a critical analysis of suggested
applications should be undertaken prior to its commencement (Bennett and
MacPherson 1985). In systems where a valid model can beneficially be
constructed, a range of advantageous uses exists, including -
1. Models can be used for manipulations and experiments which would be
impractical, too expensive, too lengthy, or impossible in the real world.
When a verified, valid model has been constructed, a wide range of
experiments can be conducted at very minimal cost compared to
traditional agricultural research.
2. In the real system, complex interactions often exist. Multiple runs of a
model can be used to identify and quantify these, further justifying this
approach as an alternative to reductionist research which may only
consider one dimension of the overall problem.